Artificial Intelligence in Urban Planning Training Course

Participants can assess whether available data is fit for an AI planning use case.

• Spatial, administrative, sensor and community data sources for planning
• Data lineage recorded from collection through transformation and model use
• Completeness, timeliness, accuracy and representativeness assessed using data profiling
• Common geographic identifiers aligned across parcels, addresses, zones and wards
• Missing values, duplicates and boundary changes documented before modelling
• Data dictionaries define fields, permitted values, owners and update cycles

Participants can prepare geospatial datasets for reproducible AI analysis.

• Coordinate reference systems selected for distance, area and overlay accuracy
• Geocoding quality checked through match rates and positional error samples
• Spatial joins connect planning applications, parcels, census areas and services
• Raster and vector formats matched to classification and prediction tasks
• Feature engineering derives density, proximity, accessibility and land-use measures
• Metadata captures source licences, processing steps and spatial limitations

Participants can define a bounded AI use case with measurable benefits and constraints.

• Theory of change links model outputs to planning service outcomes
• Baseline processes measured before proposing automation or prediction interventions
• User stories specify officers, decisions, evidence and escalation routes
• Success metrics separate model performance from policy and service outcomes
• Assumptions register records dependencies, exclusions, uncertainties and decision owners
• Prioritisation matrix compares value, feasibility, risk and implementation effort

Participants can select appropriate supervised and unsupervised learning approaches.

• Supervised learning predicts labelled outcomes from historical planning records
• Classification models sort applications, assets or locations into defined categories
• Regression estimates numeric outcomes such as demand, delay or emissions
• Clustering identifies spatial patterns without predefined outcome labels
• Training, validation and test sets prevent misleading performance claims
• Feature selection balances predictive value, explainability and data availability

Participants can interpret model performance and set acceptance thresholds.

• Confusion matrices reveal false approvals, missed risks and correct classifications
• Precision, recall and F1 scores matched to planning decision consequences
• Calibration checks whether predicted probabilities match observed outcome frequencies
• Cross-validation estimates performance stability across different data samples
• Error analysis examines geographic, demographic and temporal performance differences
• Threshold selection balances workload, harm, service targets and appeal rights

Participants can assess imagery-based methods for land and infrastructure monitoring.

• Satellite imagery resolution matched to monitoring questions and update frequency
• Image classification identifies land cover, surface change and construction activity
• Object detection locates buildings, roads, vehicles and visible infrastructure
• Change detection compares time-series imagery for unauthorised development indicators
• Ground-truth sampling validates imagery outputs against field observations
• Cloud cover, seasonal variation and occlusion documented as interpretation limits

Participants can use AI forecasts appropriately within planning scenarios.

• Time-series forecasting estimates service demand, traffic volumes and permit workloads
• Leading indicators selected from development, mobility, economic and demographic data
• Scenario assumptions distinguish forecast outputs from policy choices
• Spatial interpolation estimates conditions between observed monitoring locations
• Backtesting compares historical forecasts with actual planning and service outcomes
• Uncertainty intervals communicated alongside point estimates and maps

Participants can specify safe NLP applications for planning documents and consultation data.

• Text classification routes correspondence, applications and consultation submissions efficiently
• Topic modelling groups recurring issues within large consultation datasets
• Named entity recognition extracts locations, organisations and infrastructure references
• Sentiment analysis limitations considered for multilingual and contested public submissions
• Sampling protocols check automated coding against human reviewer judgements
• Records retention requirements applied to prompts, outputs and source texts

Participants can establish controlled uses of generative AI in planning teams.

• Approved use cases include drafting, summarisation, search and translation support
• Prompt templates require sources, uncertainty statements and prohibited assumptions
• Retrieval-augmented generation grounds outputs in approved planning policy documents
• Human review checks factual accuracy, legal interpretation and planning relevance
• Sensitive data controls prevent unauthorised disclosure through public AI services
• Version records preserve prompts, source material, outputs and reviewer decisions

Participants can identify and mitigate unequal effects from AI-supported planning decisions.

• Equality impact assessment identifies groups affected by automated recommendations
• Proxy variables examined for indirect discrimination within spatial and demographic data
• Disaggregated metrics compare error rates across relevant population groups
• Historic enforcement patterns tested before training compliance or inspection models
• Accessibility requirements considered for digital interfaces and automated communications
• Mitigation options include redesign, human review, restricted use and withdrawal

Participants can define lawful controls for personal and sensitive planning data.

• Data protection impact assessments structure privacy risks and mitigation actions
• Purpose limitation restricts reuse of planning and resident information
• Data minimisation removes fields unnecessary for the stated model purpose
• Access controls separate data preparation, model administration and decision approval
• Retention schedules govern training data, outputs, logs and model versions
• Security testing addresses data poisoning, prompt injection and unauthorised access

Participants can design governance arrangements for accountable AI planning services.

• NIST AI RMF functions organise governance, mapping, measurement and management
• ISO/IEC 42001 roles support an AI management system approach
• Algorithmic impact assessments document purpose, affected groups and safeguards
• Model cards explain intended use, data, performance and known limitations
• Human oversight protocols define intervention points, authority and escalation
• Audit trails link inputs, model versions, recommendations and final decisions

Participants can write procurement requirements that protect public control and auditability.

• Requirements specify decision purpose, users, exclusions and acceptance criteria
• Data ownership clauses cover reuse, portability, deletion and subcontractors
• Supplier evidence requested for training data, testing and model limitations
• Acceptance testing includes accuracy, fairness, security and workflow performance
• Contractual audit rights permit independent review of systems and records
• Exit plans address data return, service continuity and model decommissioning

Participants can plan an AI implementation that fits planning operations and staff roles.

• Service blueprints map resident journeys, officer tasks and system handoffs
• Operating procedures define review queues, exceptions and escalation routes
• Pilot design limits scope while collecting evidence for scaling decisions
• Training needs assessed for planners, analysts, managers and front-line staff
• Change communications explain purpose, safeguards, limitations and staff responsibilities
• Benefits realisation plans track outcomes, costs, adoption and unintended effects

Participants can communicate AI use transparently to affected communities and decision-makers.

• Stakeholder mapping identifies residents, businesses, advocates, officials and regulators
• Plain-language notices explain AI purpose, data use and human involvement
• Participation plans include accessible channels for feedback, challenge and appeal
• Public reporting presents performance, incidents, safeguards and improvement actions
• Risk communication distinguishes evidence, uncertainty, prediction and final decisions
• Community concerns translated into design requirements and governance controls

Participants can operate a monitoring regime for deployed AI planning systems.

• Key performance indicators track accuracy, service outcomes, equity and adoption
• Data drift monitoring detects changing patterns within operational planning data
• Model drift reviews identify declining performance after deployment
• Incident management records harms, near misses, complaints and corrective actions
• Periodic audits test documentation, controls, performance and human oversight
• Retirement criteria define when systems require replacement, suspension or closure

Participants can build, review and present a complete governed AI urban planning project pack.

• Problem definition links a planning decision to measurable public value
• Data inventory records sources, quality, ownership, permissions and limitations
• Method selection justifies model approach, baseline and evaluation measures
• Governance design specifies oversight, privacy, fairness, records and accountability
• Procurement requirements define supplier evidence, acceptance tests and audit rights
• Implementation roadmap sequences pilot, review gates, deployment and ongoing monitoring